Portfolio
Faizan.
AI Engineer  ·  Builder  ·  Researcher
Faizan Iqbal
Available · CSE Student · India
Faizan.

Building AI systems, full-stack products and research that solve real problems.

From building the first Kashmiri LLM to shipping full-stack products with real users — I solve hard problems at the intersection of AI and software engineering.

CS Engineering student · AI Engineer · Full-Stack Builder

View Projects → Download Resume ↓
📍 India 🎓 CSE Student ⚡ Open to Opportunities
Featured on Brut India
faizan ~ /profile
$cat profile.json
{
"flagship":"KoshurAI — First Kashmiri LLM",
"research":"3 arXiv preprints · cs.CL · cs.CV · cs.AI",
"products":[
"Groopik — Real-time Media Platform",
"PromptForge — AI Prompt Agent"
],
"featured_on":"Brut India",
"status":"open_to_work"
}
$
Scroll
// 01 — Featured Projects

Work that matters.

Not side projects. Real systems built to solve real problems at scale.

1 · Language AI
Project 01 · Language AI · Research
KoshurAI
Specialized Neural Machine Translation & Language Infrastructure for Kashmiri

7 million+ people speak Kashmiri — yet no large language model existed for the language. KoshurAI is the first Kashmiri LLM, built from scratch with a 150K+ human-verified translation pair dataset and fine-tuned on 2M+ tokens. Presented at AI Impact Summit 2026.

The Problem

Kashmiri is a low-resource language spoken by 7M+ people with near-zero representation in modern AI systems. No translation datasets, no NLP tools, no LLM — leaving an entire linguistic community excluded from the AI revolution.

Key Metrics
150K+
Translation Pairs
2M+
Language Tokens
7M+
People Affected
Training Pipeline
Data Collection→ Dataset Curation→ Tokenization→ LoRA Fine-tuning→ QLoRA Optimization→ Evaluation
Tech Stack
PyTorchHugging Face LoRAQLoRA CUDATransformers PythonDatasets API
Model on HuggingFace ↗ GitHub →
Project 02 · Language AI · Published Research
Koshur Diacritizer
Byte-Level Diacritic Restoration for Kashmiri · arXiv:2606.15883

Kashmiri text frequently omits diacritics, creating lexical ambiguity that breaks every downstream NLP task. Koshur Diacritizer is a ByT5-small seq2seq model that restores them — published on arXiv, released on Hugging Face, and covered by ETV Bharat.

The Problem

The same undiacritized Kashmiri surface form can map to several distinct words. Tokenizers, translation systems, search, and TTS all degrade without diacritics — and no neural solution existed for Kashmiri.

Key Metrics
0.0212
DERm
77.5%
Expert Accuracy
23.7K
Dataset Pairs Released
Model Pipeline
Undiacritized Text→ UTF-8 Byte Encoding→ ByT5 Encoder→ ByT5 Decoder→ Diacritized Output
Tech Stack
ByT5-smallPyTorch Hugging FaceSeq2Seq jiwer · DERm · WER
Paper on arXiv ↗ Model on HF → Press · ETV Bharat →
2 · Vision AI
Project 03 · Vision AI · Research Program
KoshurOCR
5-Generation OCR Research Program for Kashmiri Nastaliq

Five architectural generations across ~11 experimental iterations on a ~673K-sample training ecosystem — the journey from a CNN baseline to foundation-model OCR, and the discovery that re-pointed the whole program.

The Journey → The Finding

Each generation eliminated an architecture family: CNNs lacked sequence understanding, CTC plateaued on ligatures, PARSeq failed to generalize to Nastaliq, and TrOCR hit tokenizer walls. v5 (Arabic-TrOCR fine-tune) works best — and exposed decoder sequence collapse: the encoder learns Kashmiri glyphs; the decoder truncates long sequences. The bottleneck is no longer vision, it's decoding.

Key Metrics
5
Architecture Generations
~11
Experimental Iterations
673K
Training Samples
Research Timeline
v1 · CNN→ v2 · CRNN + CTC→ v3 · PARSeq→ v4 · TrOCR→ v5 · Arabic-TrOCR
Tech Stack
PyTorchTrOCR ViTBiLSTM · CTC OpenCVHugging Face
View on GitHub ↗ Dataset on HF → ⚡ v5 Active — decoder-collapse remediation
Project 04 · Vision AI · Published Dataset
Koshur Pixel
Large-Scale Synthetic OCR Dataset · arXiv:2606.23144

The first large-scale synthetic OCR dataset for Kashmiri Nastaliq — 613,078 image-text pairs generated with the SynthOCR-Gen framework, published on arXiv. The data-side intervention that powers the KoshurOCR research program.

The Problem

OCR for low-resource scripts is gated by data, not architecture. Manual annotation of Nastaliq doesn't scale — contextual ligatures make character-level labeling expensive and error-prone. Synthetic generation is the only path to scale.

Key Metrics
613K
Image-Text Pairs
25+
Augmentation Strategies
3
Granularities · Word / Line / Page
Generation Pipeline
Corpus Sampling→ Font Rendering→ Layout Engine→ Augmentation→ Validation
Tech Stack
SynthOCR-GenPython Pillow · OpenCVfonttools KS-PRET-5M corpus
Paper on arXiv ↗ GitHub →
3 · Evaluation & Benchmarking
Project 05 · Evaluation · Under Review
PhysicalHazardBenchmark
Benchmarking LLMs on Multi-Sensor Physical Hazard Assessment

Can frontier LLMs read sensor data and reason correctly about physical danger? A 68-scenario benchmark evaluating 5 frontier models across toxic gas, thermal, mechanical, intrusion, and deliberately ambiguous edge cases. Paper under review.

The Problem

LLMs are being wired into industrial monitoring, smart buildings, and robotic safety systems — but no public benchmark tests whether they can turn multi-sensor readings into correct hazard assessments and appropriate actions. Existing benchmarks test academic reasoning, not physical-world judgment.

Key Metrics
68
Scenarios
5
Frontier LLMs Evaluated
4
Evaluation Axes
Evaluation Pipeline
Sensor Context→ LLM Under Test→ Hazard Classification→ Action Check→ Coherence Score
Tech Stack
PythonBenchmark Design 5 provider APIsStructured rubrics
View on GitHub ↗ 📄 arXiv · Under Review
4 · AI Systems
Project 06 · AI Systems · Live Product
Groopik
Real-time Media Sharing Platform

A scalable real-time media sharing platform built around instant media uploads, room-based sharing, and live collaboration. Architected with Supabase for real-time sync and Cloudflare R2 for globally distributed media delivery.

The Problem

Existing media sharing tools are bloated, slow, and not built for real-time group contexts. Groopik is purpose-built for instant, room-based collaborative media sharing with zero friction.

Key Decisions
R2
Cloudflare Storage
RT
Real-time Sync
∞
Scalable Rooms
Architecture
Next.js Client→ Supabase Realtime→ Cloudflare R2→ CDN Delivery
Tech Stack
Next.jsSupabase Cloudflare R2TypeScript PostgreSQLTailwind
Live — groopik.com ↗ GitHub →
Project 07 · AI Systems · Product
PromptForge
AI Prompt Engineering Assistant

A voice-first AI assistant that helps users craft precision prompts through iterative clarification. PromptForge guides users from vague intent to optimized, structured prompts — reducing LLM output variance and improving task completion rates.

The Problem

Most users get poor results from LLMs not because the model is bad — but because their prompts are imprecise. PromptForge solves prompt quality at the input layer through voice and structured clarification.

Core Features
🎙️
Voice Input
🔄
Clarification Loop
⚡
Prompt Optimizer
Workflow
Voice / Text Input→ Intent Analysis→ Clarification→ Prompt Generation→ Output
Tech Stack
ReactGroq · Llama 3.3 70B Anthropic ClaudeWeb Speech API Vercel FunctionsTypeScript
Live Demo ↗ GitHub →
// 02 — Research

Published work.

3 arXiv preprints and an ongoing multi-generation OCR research program — spanning NLP, computer vision, and LLM evaluation.

Published arXiv:2606.15883 · cs.CL · cs.AI

Koshur Diacritizer: A Byte-Level Sequence-to-Sequence Model for Kashmiri Diacritic Restoration

ByT5-small model for restoring diacritics in Kashmiri text — DERm 0.0212, WER 0.2159, 77.5% expert accuracy. Includes the first public dataset of 23.7K aligned diacritized Kashmiri sentence pairs. Covered by ETV Bharat.

View on arXiv ↗   Model on HuggingFace ↗   Press · ETV Bharat ↗
Published arXiv:2606.23144 · cs.CV · cs.CL

Koshur Pixel: A Large-Scale Synthetic OCR Dataset for Kashmiri

613,078 image-text pairs for Kashmiri OCR, generated via SynthOCR-Gen from the KS-PRET-5M corpus. First large-scale synthetic OCR dataset for the Kashmiri script, covering 25+ augmentation strategies across multiple fonts and granularities.

View on arXiv ↗   View on GitHub ↗
Under Review arXiv · cs.AI

Benchmarking LLMs on Multi-Sensor Physical Hazard Assessment

68-scenario benchmark evaluating 5 frontier models on physical hazard reasoning across multi-sensor contexts.

View on GitHub ↗
Ongoing Program 5 generations · ~673K samples

KoshurOCR: Multi-Generation OCR Research Program for Kashmiri Nastaliq

Five architectural generations (CNN → CRNN+CTC → PARSeq → TrOCR → Arabic-TrOCR) across ~11 experimental iterations on a ~673K-sample training ecosystem. Key finding: decoder sequence collapse under heterogeneous supervision — the bottleneck is no longer visual representation learning but robust long-sequence autoregressive decoding. Released Koshur-OCR-Synth, a solo-built 571,740-pair clean synthetic OCR dataset.

View on GitHub ↗   Dataset on HuggingFace ↗
Datasets Released · Public
613K
Synthetic OCR Pairs · Koshur Pixel (co-authored)
571K
Clean OCR Pairs · Koshur-OCR-Synth (solo)
23.7K
Diacritized Sentence Pairs
150K+
Translation Pairs
Kashmiri–English dataset
2M+
Language Tokens
KoshurAI training corpus
0.698
COMET Score
KS→EN · +0.0241 vs Meta NLLB-200
3
arXiv Preprints
cs.CL · cs.CV · cs.AI
673K+
OCR Training Samples
KoshurOCR + Koshur Pixel ecosystem
3×
State Champion
J&K Wushu Championship
// 03 — Achievements

Built to compete.

AI summits, national podiums, state championships. Each achievement is proof, not decoration.

2026 · Research
2 arXiv Papers Published · 1 Under Review
Koshur Diacritizer + Koshur Pixel published · ETV Bharat national coverage
📄 arXiv
2025 · Robotics
Rank #1 — IIT Mandi Xpecto'25
1000+ participants · Beat senior engineers from top institutes
🥇 National Winner
2025 · Robotics
3rd Place — IIT Jodhpur Prometeo'25
National robotics combat · Built the bot. Competed. Placed.
🥉 National Podium
2024 · Robotics
Winner — RC-Xtreme 3.0 Robotics Competition
Remote-controlled systems · Speed, precision, engineering
🏆 Champion
2026 · AI Research
University Representative — AI Impact Summit 2026
Presented 15 AI projects · Featured on Brut India · National coverage
🤖 AI Research
2025 · Recognition
Met Chief Minister of Punjab
Startup Punjab Conclave 2025 · Rooms most students never enter
Ongoing · Leadership
Outreach Head — RISC Robotics Club, LPU
Coordinated with Indian Army · 30+ certificates from IITs, NITs, Corporates
2022–2024 · Athletics
3× J&K Wushu State Champion
Martial arts · Discipline, physical consistency · Three consecutive titles
🥊 3× Champion
// 04 — Tech Stack

Tools of the trade.

From LLM fine-tuning to full-stack product engineering.

🔥PyTorch
🤗Hugging Face
🔄LoRA
⚡QLoRA
🖥️CUDA
🧠Transformers
📊Datasets API
🐍Python
📈Weights & Biases
💎Gemma 3
🔤ByT5
👁️TrOCR / ViT
🎯COMET
🌐FLORES-200
⚛️React
▲Next.js
🔷TypeScript
💨Tailwind CSS
🎨CSS / Sass
📦Vite
🌐HTML5
🟢Node.js
🚂Express
🐘PHP
🔌REST APIs
🔑JWT / Auth
🧪Postman
🍃MongoDB
🐘PostgreSQL
⚡Supabase
🔴Redis
📊SQL
☁️AWS
🌩️Google Cloud
🐳Docker
▲Vercel
🔶Cloudflare R2
🐙GitHub Actions
🖥️Linux
🤖Arduino
👁️OpenCV
⚙️Embedded C
📡IoT Protocols
🔌Raspberry Pi
🛠️Circuit Design
// 05 — Current Focus

What's being built.

Active engineering threads right now.

👁️

KoshurOCR v5

Remediating decoder sequence collapse in Nastaliq OCR — length-aware decoding objectives and short→long curriculum learning over the 673K-sample ecosystem.

Active
🎙️

Koshur Knowledge Assistant

Speech + RAG over the Kashmiri corpus — the next pillar of the Kashmiri AI stack, building on KoshurAI, the Diacritizer, and OCR.

🔬 Prototyping
🌱

Open Source & Products

Contributing to AI/agent projects via GirlScript Summer of Code 2026, while iterating on Groopik and PromptForge with real user feedback.

Shipping
// 06 — About

From Kashmir,
building for the world.

I'm a builder from Kashmir who enjoys solving difficult problems through AI, full-stack software, and product engineering. I don't believe in building things for the sake of building — every project I take on has a clear problem statement and a measurable outcome.

From creating the first Kashmiri LLM to shipping full-stack products with real users and publishing AI research, I focus on work that creates measurable impact for real people.

I'm equally at home training a model on CUDA clusters, designing a PCB for a combat robot, or architecting a scalable backend. That cross-domain fluency is what I bring to every team.

Interests
  • 🧠AI Research & LLM Fine-Tuning
  • 🤖Robotics & Embedded Systems
  • 🌍Low-Resource & Multilingual AI
  • ⚙️Product Development
  • 🏗️Systems Architecture

// Profile

LocationIndia 🇮🇳
ProgramCSE (B.Tech)
UniversityLPU
FocusAI + Full-Stack
StatusOpen to Work ✓
Emailifaizan041@gmail.com
// 07 — Beyond Software

Same engine.
Different arena.

The discipline that wins Wushu championships is the same discipline that ships code at 2am.

🥊

3× Wushu Champion

J&K State Level

🤖

Robotics Competitor

National IIT Podiums

🎤

Public Speaker

AI Summits & Events

🔬

Research Presenter

15 AI Projects Shown

🏹

Outdoor Pursuits

Archery · Football · Swim

// 08 — Recognition

Work that gets noticed.

Press, events, and stages that created credibility beyond the code.

2026

National Media Feature

Featured by Brut India as university representative at AI Impact Summit 2026 — presenting 15 AI projects. National digital coverage reaching millions.

2026

University Representative

Selected as university representative to present 15 AI projects at national AI Impact Summit 2026.

2025

Startup Punjab Conclave 2025

Met Chief Minister of Punjab at the Startup Punjab Conclave — rooms most students never enter.

2025

Xpecto'25 — Rank #1

Top-ranked competitor at IIT Mandi's Xpecto'25 robotics championship among 1000+ participants.

2025

Prometeo'25 — 3rd Place

National podium finish at IIT Jodhpur's Prometeo'25 robotics combat competition.

2026

National Press — Kashmiri AI Research

"Three Kashmir Engineers Develop AI Tool to Preserve Native Language in Digital Age" — national coverage of our Kashmiri language AI work (KoshurAI · Koshur Diacritizer).

Ongoing

RISC Robotics Club Collaboration

Coordinated robotics outreach with Indian Army through RISC club as Outreach Head, LPU.

// 09 — Certifications

Verified credentials.

30+ certificates from globally recognized platforms. A selection of the most relevant.

NPTEL
Privacy & Security in Online Social Media
Nov 2025
Verify ↗
LPU
Data Structures and Algorithms
July 2025
Verify ↗
Coursera
Fundamentals of Network Communication
Nov 2024
Verify ↗
IBM
Introduction to Hardware and Operating Systems
Sept 2024
Verify ↗
FreeCodeCamp
Responsive Web Design
Sept 2023
Verify ↗
GLA University
Data Science Foundations
Aug 2023
Verify ↗
// 10 — Contact

Building something
ambitious?

Let's talk.